Novel Phenotyping for Acute Heart Failure-Unsupervised Machine Learning-Based Approach.

Area of Science:

• Cardiology
• Medical Data Science
• Machine Learning in Healthcare

Background:

• Acute heart failure (AHF) presents as a heterogeneous and life-threatening condition.
• Clinical presentation and management of AHF are influenced by etiological factors, cardiac substrate, and comorbidities.

Purpose of the Study:

• To analyze the inherent phenotypic heterogeneity within the AHF patient population.
• To assess the utility of unsupervised machine learning (clustering) for medical data analysis in AHF.

Main Methods:

• Utilized K-medoids clustering on data from 381 AHF patients.
• Included 63 clinical and biochemical features assessed at patient admission.
• Optimized clustering using the Davies-Bouldin index, blinded to patient outcomes.

Main Results:

• The algorithm identified six distinct clusters (phenotypes) among AHF patients.
• These clusters differed significantly across 58 variables, including etiology, clinical status, comorbidities, and laboratory parameters.
• Significant differences in one-year mortality were observed between the identified clusters (p = 0.002).

Conclusions:

• Clustering successfully extracted six distinct AHF phenotypes with unique clinical characteristics and prognoses.
• These findings provide a foundation for future clinical trials and the development of tailored AHF treatment approaches.